Drug discovery using very large numbers of patents. General strategy with extensive use of match and edit operations

Abstract

A patent data base of 6.7 million compounds generated by a very high performance computer (Blue Gene) requires new techniques for exploitation when extensive use of chemical similarity is involved. Such exploitation includes the taxonomic classification of chemical themes, and data mining to assess mutual information between themes and companies. Importantly, we also launch candidates that evolve by “natural selection” as failure of partial match against the patent data base and their ability to bind to the protein target appropriately, by simulation on Blue Gene. An unusual feature of our method is that algorithms and workflows rely on dynamic interaction between match-and-edit instructions, which in practice are regular expressions. Similarity testing by these uses SMILES strings and, less frequently, graph or connectivity representations. Examining how this performs in high throughput, we note that chemical similarity and novelty are human concepts that largely have meaning by utility in specific contexts. For some purposes, mutual information involving chemical themes might be a better concept.

Appendix: Binding studies

Binding studies will be more familiar, but are well known to be nontrivial. Note first that in the catalytic region of chain A of the experimental 2BEL tetramer structure with carbenoxolone and NADP cofactor ligands (and rather similarly for the other monomers), the serine Ser 170 side chain oxygen atom makes a surprisingly tight 2.5 Ǻ approach to the original double bonded C=O oxygen O11 atom on C11 of the steroid-like framework. The sum of the oxygen atom van der Waal’s radii of the serine and ligand oxygen is 1.52 + 1.52 = 3.04 Ǻ. Following DOCK and AMBER, this O…O distance is 3.0 Ǻ. Adding to the electronegative tension, there is also a close approach by the phenolic oxygen of Tyr 183 at 5.0 Ǻ, and by oxygen O7 N in the NADP cofactor at 4.9 Ǻ. The other close approaches do not compensate significantly: they are most importantly the carbons C4 N of NADP at 4.3 Ǻ and the Ala 172 side chain carbon at 4.5 Ǻ, but a good binding comes from interactions involving the whole binding site region. It is substitutions at or around C11 that are compensated by the most surprising degree of binding site accommodation, i.e. significant conformational changes in backbone as well as side-chains in the binding site. Replacing the O…O contact by O…S by using thioketone (cboS1) as ligand initially generates a considerable van der Waal’s and electrostatic repulsion as the thioketone S is strongly electronegative [49], but binding is still accommodated. These observations led to the following studies.

In order (a) to see if any multiple binding modes and binding site conformational changes may have been missed, and (b) to try and predict Blue Gene performance as a screening tool for potential candidates, a variety of extensive analyses and comparative studies using the full tetramer with NADP cofactors were performed using KRUNCH [50]. This can use and compare AMBER and other force fields and re-parameterize them, as well as perform both molecular mechanics and dynamics calculations. As an alternative to Blue Gene’s processing speed and brute force calculation, this has a large kit of techniques for spanning energy barriers and searching conformational space, developed and adapted over many years by Robson and coworkers. See Ref. [51] for a general review, and techniques of Refs. [52–54] were also used. Results relative to carbenoxolone are shown in column 3 of Table 2. The regression slope of columns 2 and 3 is close to unity at 0.97 with a Pearson’s correlation R of 0.88, with intercept close to zero. That is after adjusting slope for predictive purposes: prior to the latter, on average the RPFF energies were amplified 4.6% over the DOCK-AMBER results. The major difference here appears to be in electrostatics, because a simple 6% increase in effective dielectric constant in the RPFF brings the results into alignment (regression slope 0.95). Even recalling that this is predicting another calculation, not experiment, the above is a reasonable result within the state of the art. However, in more recent studies, probes started from the patent and ZINC data bases were left to evolve several weeks. Some ligands bound specifically to the 2BEL binding site by the RPFF plus heuristic methods, but did not bind to the site using DOCK and AMBER. An early example was (2,3)-dithio-6-hydroxy-8-carboxy-(1,7,9)-azanaphthalene as indicated by ‘**’ in the second column of Table 2. The common feature of these “prediction failures” is that they are small ligands of one or two rings, and display multiple binding modes.